Resource Theory of Athermality and Quantum Thermodynamics

Thermodynamics was built for steam engines — objects with Avogadro's number of molecules, where fluctuations are invisible and a single temperature, entropy, and free energy suffice. Quantum information theory asks the opposite question: what survives when the working body is a handful of qubits, possibly coherent, coupled to one heat bath? The resource-theoretic answer is to stop postulating the laws of thermodynamics and start deriving them, borrowing the scaffolding entanglement theory made famous: declare some operations physically free, identify the free states you can make from nothing, and ask which transformations they permit. Fix the free resource to be a single heat bath at inverse temperature $\beta=1/k_BT$ . Then the free states are the Gibbs states $\gamma=e^{-\beta H}/Z$ , the free operations are the thermal operations — energy-conserving unitaries with thermal ancillas — and everything else is athermal: a resource. From exactly two axioms — strict energy conservation (the first law, in seed form) and a bath that starts in equilibrium (the second) — a whole edifice follows as theorems, not assumptions. The Gibbs state is a fixed point of every thermal operation, so the quantum relative entropy $D(\rho\|\gamma)$ can only decrease, and that single inequality is the second law: it says the nonequilibrium free energy $F_\beta(\rho)=\langle H\rangle_\rho-k_BT\,S(\rho)$ never rises. In the many-copy limit that one number governs everything; drop to a single copy and the law shatters into a continuum — a whole Rényi family of free energies that must each decrease — while coherence between energy levels emerges as a second, independent resource the bath can never supply.

Key takeaways

Athermality is a resource theory: one heat bath at inverse temperature $\beta$ is free, the Gibbs states $\gamma=e^{-\beta H}/Z$ are the free states, and thermal operations — energy-conserving unitaries coupled to thermal ancillas — are the free operations.
Any state that is not the Gibbs state of its Hamiltonian carries athermality: out-of-equilibrium order that can be spent as work or converted into other states.
One number governs the macroscopic law: the nonequilibrium free energy $F_\beta(\rho)=\langle H\rangle_\rho-k_BT\,S(\rho)$ can only decrease, because $D(\rho\|\gamma)=\beta[F_\beta(\rho)-F_\beta(\gamma)]$ is a data-processing monotone.
At the single-shot level the second law splits into many: the $\alpha$ -Rényi divergences $D_\alpha(\rho\|\gamma)$ must all decrease, with $F_{\max}$ (from $D_0$ ) setting distillable work and $F_{\min}$ (from $D_\infty$ ) the work of formation.
Landauer erasure, Maxwell's demon, quantum batteries, and coherence-as-a-resource all live inside this framework — the cleanest information-theoretic basis for quantum thermodynamics.
Key idea: fix one bath temperature; the Gibbs state is free and energy-conserving unitaries with thermal ancillas are free. Then relative entropy to the Gibbs state — equivalently nonequilibrium free energy — is a monotone, giving the second law; in the single-shot regime a whole Rényi family of free energies must all decrease.

Where this lives today

The athermality framework has moved from theorem to laboratory. Single-shot and Rényi free energies set the design targets for quantum batteries, where the $F_{\max}$ / $F_{\min}$ gap sets how much of a charged state's energy is deliverable in one use versus over many cycles. Landauer's $k_BT\ln2$ has been measured directly, first in a colloidal-particle bit by Bérut and colleagues in 2012 and since in single-electron boxes and superconducting circuits, confirming the floor to within a few percent. Trapped ions, NV centres, and superconducting qubits now realize small quantum heat engines and refrigerators bounded by exactly these monotones, and coherent-medium experiments probe the asymmetry-resource question free energy cannot answer. Open problems remain sharp: the exact conditions for coherent-state conversion, the role of catalysis, and how system–bath correlations modify the second law all sit at the active frontier.

The physical picture

Picture a landscape whose height is free energy. The Gibbs state $\gamma$ sits alone at the valley floor; every other state is up the slopes, and its altitude above the floor is exactly its relative entropy to equilibrium. A thermal operation is a ball that may roll downhill or drift along a contour but can never climb on its own — climbing would mean an equilibrium bath spontaneously created order. Work extraction is cashing in altitude; the work of formation is the price of being lifted up; and the second law is just the fact that free operations cannot raise you here. The first figure makes the landscape and its governing relation, $D(\rho\|\gamma)=\beta(F_\beta(\rho)-F_\beta(\gamma))$ , into one picture.

Figure

Free-energy landscape: the Gibbs state γ sits at the valley floor and every athermal state ρ is uphill by D(ρ‖γ) = β(F_β(ρ) − F_β(γ)). Thermal operations only roll downhill, so work out is at most −ΔF_β while forming ρ costs at least that much — the second law is that you cannot climb for free.

Two distinct resources hide inside an athermal quantum state. The first is a mismatch of populations — the energy levels are occupied in the wrong proportions relative to Boltzmann, and this classical free-energy resource is all a diagonal state carries. The second is genuinely quantum: coherence between energy eigenstates, a definite phase relation the bath cannot supply, because a system in equilibrium has no clock. Strict energy conservation makes every thermal operation time-translation covariant, so it can never create coherence — a conserved-charge-like asymmetry resource sitting on top of the free-energy one. The familiar single second law of the textbook is a thermodynamic-limit phenomenon: with many copies, averaging drowns the fluctuations and only $F_\beta$ survives, so one number decides every conversion. For a handful of qubits the law fractures — whether $\rho$ becomes $\sigma$ depends on infinitely many inequalities $D_\alpha(\rho\|\gamma)\ge D_\alpha(\sigma\|\gamma)$ , one per Rényi index $\alpha$ , each probing a different slice of the eigenvalue spectrum. The second figure shows this Rényi tower for a concrete state and how it collapses in the many-copy limit.

Figure

The tower of second laws for a qutrit with Gibbs weights (1/2, 1/4, 1/4) and populations (1/2, 1/2, 0): the generalized free energies rise with α — D_0 = 0.288, D_1 = 0.347, D_∞ = 0.693 nats — and every one must decrease under a thermal operation. In the many-copy i.i.d. limit they collapse onto D_1, the single macroscopic second law.

The core equations

The theory is fixed by naming its free states and free operations. The free state is the Gibbs state — the unique state a lone bath at temperature $T$ relaxes to.

\gamma=\frac{e^{-\beta H}}{Z},\qquad Z=\mathrm{Tr}\,e^{-\beta H},\qquad \beta=\frac{1}{k_BT} \qquad (1)

where: $H$ — the system Hamiltonian; $\beta$ — inverse temperature of the single fixed bath; $Z$ — partition function that normalizes the state; $k_B,\;T$ — Boltzmann's constant and the bath temperature. The Gibbs state is the only free state: at fixed $\beta$ it is what you get for nothing, and everything else must be paid for. A thermal operation is the most general dynamics you can build from that free resource: borrow a bath in equilibrium, couple it with a unitary that spends no external work, and discard the bath.

\mathcal{E}(\rho)=\mathrm{Tr}_B\!\big[\,U\,(\rho\otimes\gamma_B)\,U^\dagger\big],\qquad [\,U,\,H_S+H_B\,]=0 \qquad (2)

where: $\gamma_B$ — ancilla bath in its own Gibbs state $e^{-\beta H_B}/Z_B$ ; $U$ — global unitary that strictly conserves total energy; $\mathrm{Tr}_B$ — partial trace discarding the bath after the interaction. No work reservoir is hidden inside $U$ , so any work in a real protocol must be booked explicitly on a separate battery. Each state's resource is its nonequilibrium free energy — the textbook $F=U-TS$ , but valid for any state, not only equilibrium ones.

F_\beta(\rho)=\langle H\rangle_\rho-k_BT\,S(\rho)=\mathrm{Tr}(H\rho)-k_BT\,S(\rho) \qquad (3)

where: $\langle H\rangle_\rho=\mathrm{Tr}(H\rho)$ — the average energy; $S(\rho)=-\mathrm{Tr}(\rho\log\rho)$ — the von Neumann entropy. For $\rho=\gamma$ this reduces to the equilibrium free energy $F_\beta(\gamma)=-k_BT\log Z$ , the valley floor of the landscape. The bridge that makes the theory tractable is that this free energy is nothing but relative entropy to the Gibbs state, up to a constant — distance from equilibrium and free energy are the same quantity in two languages.

D(\rho\|\gamma)=\mathrm{Tr}\big[\rho\log\rho-\rho\log\gamma\big]=\beta\big[F_\beta(\rho)-F_\beta(\gamma)\big] \qquad (4)

where: $D(\rho\|\gamma)\ge0$ — quantum relative entropy of $\rho$ to $\gamma$ , zero only at equilibrium; $F_\beta(\gamma)=-k_BT\log Z$ — the constant offset. An athermal state is literally "far" from $\gamma$ , and that distance is its stored free energy in units of $k_BT$ . Because $\gamma$ is a fixed point of every thermal operation and relative entropy contracts under all quantum channels, $D$ — hence $F_\beta$ — can only fall: the macroscopic second law, derived rather than assumed.

\rho\xrightarrow{\text{thermal op}}\sigma\ \Longrightarrow\ D(\rho\|\gamma)\ge D(\sigma\|\gamma)\ \Longleftrightarrow\ F_\beta(\rho)\ge F_\beta(\sigma) \qquad (5)

where: the arrow means "reachable by some thermal operation." This is exactly Clausius's statement that free energy cannot spontaneously increase. For a single copy that inequality is not enough: the exact conditions call on a whole family of generalized free energies built from the Rényi divergences, each an independent monotone.

D_\alpha(\rho\|\gamma)=\frac{1}{\alpha-1}\log\mathrm{Tr}\big[\rho^{\alpha}\gamma^{1-\alpha}\big],\qquad F_\alpha(\rho)=k_BT\,D_\alpha(\rho\|\gamma)+F_\beta(\gamma) \qquad (6)

where: $\alpha\ge0$ — the Rényi index; for states commuting with $H$ each $\alpha$ gives a separate second law. The endpoints matter most: $D_0=-\log\mathrm{Tr}[\Pi_\rho\gamma]$ (the max-divergence, giving $F_{\max}$ ) and $D_\infty=\log\max_i(\lambda_i/\gamma_i)$ (the min-divergence, giving $F_{\min}$ ), and they obey $D_0\le D\le D_\infty$ . Finally the operational payoff: work is bounded by the free-energy drop, and in the trivial-Hamiltonian corner this becomes Landauer's principle.

\langle W_{\mathrm{ext}}\rangle\le -\Delta F_\beta=k_BT\big[D(\rho\|\gamma)-D(\rho'\|\gamma)\big],\qquad W_{\mathrm{erase}}\ge k_BT\ln2 \qquad (7)

where: $W_{\mathrm{ext}}$ — average work drawn into an explicit battery; $\rho'$ — the final system state; $W_{\mathrm{erase}}$ — the work to reset one bit to a pure state when $H=0$ . You extract at most the free energy you spend, and even to forget one bit you pay $k_BT\ln2$ .

Derivation ARelative entropy to the Gibbs state is a thermal-operation monotone

The goal is to show that the entire macroscopic second law follows from two facts that are each easy to establish alone: that the Gibbs state is a fixed point of every thermal operation, and that quantum relative entropy can never increase under any completely positive trace-preserving map. Chaining them makes $D(\rho\|\gamma)$ fall under thermal operations, and equation (4) turns that into the statement that free energy cannot rise — Clausius's law, obtained without ever postulating it.

Step 1. Establish the fixed point $\mathcal{E}(\gamma)=\gamma$ . The uncoupled thermal product $\gamma_S\otimes\gamma_B\propto e^{-\beta(H_S+H_B)}$ is a function of $H_S+H_B$ , and $U$ commutes with it, so $U(\gamma_S\otimes\gamma_B)U^\dagger$ is unchanged; tracing out $B$ returns $\gamma_S$ . Why: a bath at equilibrium plus energy conservation cannot push the system out of equilibrium.

Step 2. Invoke data processing. Quantum relative entropy obeys $D(\mathcal{N}(\rho)\|\mathcal{N}(\sigma))\le D(\rho\|\sigma)$ for every CPTP map $\mathcal{N}$ — Lindblad–Uhlmann monotonicity, equivalent to strong subadditivity. A thermal operation is CPTP, so $D(\mathcal{E}(\rho)\|\mathcal{E}(\gamma))\le D(\rho\|\gamma)$ . Why: distinguishability from $\gamma$ can only be lost under a noisy channel, never manufactured.

Step 3. Substitute the fixed point from Step 1: $\mathcal{E}(\gamma)=\gamma$ turns the left side into $D(\mathcal{E}(\rho)\|\gamma)\le D(\rho\|\gamma)$ . Why: replacing the evolved reference by the untouched Gibbs state is what makes $D(\cdot\|\gamma)$ a genuine monotone.

Step 4. Translate to free energy with equation (4). Since $D(\rho\|\gamma)=\beta[F_\beta(\rho)-F_\beta(\gamma)]$ with $F_\beta(\gamma)$ constant, the inequality reads $F_\beta(\mathcal{E}(\rho))\le F_\beta(\rho)$ . Why: this is the physical second law — free energy is non-increasing under every thermal operation. *Verified: the two ingredients — Gibbs invariance from $[U,H_S+H_B]=0$ and data processing (Wilde, Quantum Information Theory, 2nd ed.) — are independent and each checkable; together they reproduce the Clausius inequality.*

Go deeper · Derivation B · 1 minAverage extractable work is capped by the free-energy decrease

Derivation BAverage extractable work is capped by the free-energy decrease

The aim is to turn the abstract monotone into a hard engineering bound: couple the system to an explicit work reservoir — a weight or battery whose energy we can read off — and show that energy conservation plus the free-energy monotone cap the average work delivered at $-\Delta F_\beta$ , with equality only for reversible, quasi-static protocols. This is the quantum maximum-work theorem, and its trivial-Hamiltonian corner is Landauer's principle.

Step 1. Book the energy. Run a global thermal operation on system, bath, and weight together. Total energy is conserved, so the work delivered to the weight is $\langle W_{\mathrm{ext}}\rangle=-\Delta E_S-\Delta E_B$ , with $\Delta E_B=Q$ the heat into the bath. Why: because $U$ conserves $H_S+H_B+H_W$ , every joule the battery gains is one the system-plus-bath lost.

Step 2. Impose the second law on the system's free energy: $F_\beta(S)$ cannot increase, so the recoverable work is bounded by its decrease, $-\Delta F_\beta(S)=-\Delta E_S+k_BT\,\Delta S(S)\ge0$ . Why: the monotone of the previous derivation is the resource whose depletion pays for the work.

Step 3. Use the Clausius bound $k_BT\,\Delta S(S)\ge -Q$ : the bath can only absorb heat consistent with its entropy rising. Why: this is where irreversibility enters — heat dumped to the bath at $T$ is entropy that cannot be reclaimed as work.

Step 4. Combine Steps 1–3: $\langle W_{\mathrm{ext}}\rangle=-\Delta E_S-Q\le-\Delta E_S+k_BT\,\Delta S(S)=-\Delta F_\beta$ , with equality for reversible thermalizations. Why: this recovers $W\le-\Delta F$ , and setting $H=0$ makes erasing one bit — a $\ln2$ drop in entropy — cost $k_BT\ln2$ , which is Landauer. *Verified: reduces to the classical maximum-work theorem; the $H=0$ special case yields $k_BT\ln2$ per erased bit, numerically $2.87\times10^{-21}\ \mathrm{J}\approx17.9\ \mathrm{meV}$ at $300\ \mathrm{K}$ .*

Go deeper · Derivation C · 1 minSingle-shot distillable work and the irreversibility gap

Derivation CSingle-shot distillable work and the irreversibility gap

The purpose here is to expose why one free energy is not enough for a single copy. Deterministic single-shot work extraction — work you are guaranteed, not merely averaged — is governed not by $F_\beta$ but by the max-free energy $F_{\max}$ from the Rényi-0 divergence, while the cost to form the same state is set by the min-free energy $F_{\min}$ from Rényi- $\infty$ . The ordering $D_0\le D\le D_\infty$ opens a positive gap between what you pay and what you get back — a single-shot irreversibility that closes only in the many-copy limit.

Step 1. Define the support divergence for a state $\rho$ diagonal in the energy basis with populations $p_i$ : $D_0(\rho\|\gamma)=-\log\mathrm{Tr}[\Pi_\rho\gamma]=-\log\sum_{i:p_i>0}\gamma_i$ , with $\Pi_\rho$ projecting onto the occupied levels. Why: deterministic distillation must succeed on every populated level at once, so only the Gibbs weight under the support can be cashed in — the worst case, not the average.

Step 2. Read off the distillable work $W^{(1)}_{\mathrm{ext}}=-\Delta F_{\max}=k_BT\,D_0(\rho\|\gamma)$ and the work of formation $W^{(1)}_{\mathrm{form}}=\Delta F_{\min}=k_BT\,D_\infty(\rho\|\gamma)$ , with $D_\infty=\log\max_i(p_i/\gamma_i)$ fixed by the most over-occupied level. Why: extraction is limited by the support overlap, formation is priced by the tallest population spike — two features of one spectrum.

Step 3. Order the divergences. Rényi divergences are non-decreasing in $\alpha$ , so $D_0\le D\le D_\infty$ , hence $F_{\max}\le F_\beta\le F_{\min}$ : distillable work sits below the average bound, which sits below the formation cost. Why: the monotonicity that gives one clean average law makes the single-shot brackets straddle it, never cross it.

Step 4. Identify the irreversibility gap $W^{(1)}_{\mathrm{form}}-W^{(1)}_{\mathrm{ext}}=k_BT[D_\infty-D_0]\ge0$ . It vanishes only for many i.i.d. copies, where the quantum asymptotic equipartition property collapses every $D_\alpha$ onto $D$ . Why: reversibility is an asymptotic luxury — one copy always loses something between formation and distillation. *Verified: Horodecki–Oppenheim (Nat. Commun. 2013); the ordering $D_0\le D\le D_\infty$ follows from Rényi monotonicity in $\alpha$ and the i.i.d. collapse from the quantum AEP. For the qutrit of Worked example 2, $D_0=0.288<D=0.347<D_\infty=0.693$ nats.*

Limiting cases

The framework interpolates between regimes as you vary the number of copies, the coherence, and the Rényi index $\alpha$ — each corner recovering a law discovered historically on its own.

Case	Result	Meaning
Asymptotic / many copies	$\rho^{\otimes n}\to\sigma^{\otimes m},\ \dfrac{m}{n}\to\dfrac{F_\beta(\rho)-F_\beta(\gamma)}{F_\beta(\sigma)-F_\beta(\gamma)}$	A single free energy governs reversible conversion; this is the ordinary second law of thermodynamics.
Single-shot extraction	$W^{(1)}_{\mathrm{ext}}=-\Delta F_{\max}=k_BT\,D_0(\rho\\|\gamma)$	Deterministic yield is set by the max-free energy (Rényi-0) and can fall far below $-\Delta F_\beta$ for noisy states.
Single-shot formation	$W^{(1)}_{\mathrm{form}}=\Delta F_{\min}=k_BT\,D_\infty(\rho\\|\gamma)$	Creating $\rho$ from equilibrium costs the min-free energy (Rényi- $\infty$ ); the gap $F_{\min}-F_{\max}$ is the irreversibility.
Trivial Hamiltonian $H=0$	$\gamma=\mathbb{1}/d,\ F_\beta=-k_BT\,S(\rho)+\text{const}$	Athermality is purely entropic; work extraction becomes Landauer erasure at $k_BT\ln2$ per bit.
Classical / diagonal states	$D(\rho\\|\gamma)=\beta\langle H\rangle_\rho-S(\rho)-\log Z$	For $[\rho,H]=0$ athermality reduces to a Boltzmann free-energy resource; coherence monotones vanish.
Coherent / asymmetry sector	$[\rho,H]\ne0\Rightarrow$ free energy alone insufficient	Time-translation covariance makes inter-level coherence an independent, non-amplifiable resource.

Common trapFree energy is a complete invariant for thermal conversion

It is not. Because thermal operations conserve energy strictly, they are time-translation covariant and cannot create, rotate, or amplify coherence between energy eigenspaces. Two states can therefore share identical populations, identical entropy, and identical values of every $F_\alpha$ — yet be inconvertible, because one carries coherence the other lacks and no thermal operation can supply it. The population-based free energies are necessary but never sufficient for coherent states: a full account also needs asymmetry monotones for the coherence resource. Treating $F_\beta$ (or even the whole Rényi tower) as the last word is the classic error, and it hides the genuinely quantum half of quantum thermodynamics.

Worked examples

1 · Work from an excited qubit, and the price of forgetting

A qubit has ground energy $0$ and excited energy $E$ , coupled to a bath with $\beta E=\ln2$ , so the Gibbs populations are $\gamma=(2/3,\,1/3)$ . It starts in the pure excited state $\rho=\lvert1\rangle\langle1\rvert$ . How much work can be extracted on average, and how does this compare to erasing a bit?

Step 1. Find the resource. A pure state has $S(\rho)=0$ , so its free energy is just its energy, $F_\beta(\rho)=E$ . The equilibrium floor is $F_\beta(\gamma)=-k_BT\log Z$ with $Z=1+e^{-\beta E}=1+\tfrac12=\tfrac32$ . The stored resource is $D(\rho\|\gamma)=\beta[F_\beta(\rho)-F_\beta(\gamma)]=\beta E+\log\tfrac32=\ln2+\ln\tfrac32=\ln3$ . So $D(\rho\|\gamma)=\mathbf{1.099\ \text{nats}}$ .

Step 2. Convert to work. The average bound (7) gives $\langle W_{\mathrm{ext}}\rangle\le k_BT\,D(\rho\|\gamma)=k_BT\ln3$ . At $T=300\ \mathrm{K}$ , $k_BT=4.14\times10^{-21}\ \mathrm{J}=25.9\ \mathrm{meV}$ , so $\langle W_{\mathrm{ext}}\rangle\le\mathbf{28.4\ meV}$ per qubit, approached only by reversible thermalization.

Step 3. Compare with erasure. Setting $H=0$ , resetting one maximally mixed bit removes $\ln2$ of entropy and therefore costs $W_{\mathrm{erase}}\ge k_BT\ln2=\mathbf{17.9\ meV}$ at $300\ \mathrm{K}$ — the same ledger run backwards. Extract at most $k_BT\ln3\approx28.4$ meV from the excited qubit; pay at least $k_BT\ln2\approx17.9$ meV to erase a bit.

2 · The Rényi tower for a qutrit: distillable vs. formation work

A three-level system has energies $(0,E,E)$ with $\beta E=\ln2$ , giving unnormalized Gibbs weights $(1,\tfrac12,\tfrac12)$ , so $Z=2$ and $\gamma=(\tfrac12,\tfrac14,\tfrac14)$ . Take the diagonal state with populations $p=(\tfrac12,\tfrac12,0)$ — occupying only the lowest two levels. Compute $D_0$ , $D$ , and $D_\infty$ , and read off the single-shot work bounds and the irreversibility gap.

Step 1. Average divergence. $D(\rho\|\gamma)=\sum_i p_i\log(p_i/\gamma_i)=\tfrac12\log\!\frac{1/2}{1/2}+\tfrac12\log\!\frac{1/2}{1/4}=0+\tfrac12\ln2=\mathbf{0.347\ \text{nats}}$ — the ordinary free-energy resource $F_\beta(\rho)-F_\beta(\gamma)=0.347\,k_BT$ .

Step 2. Max-divergence (distillation). The support is levels $0$ and $1$ , so $D_0(\rho\|\gamma)=-\log(\tfrac12+\tfrac14)=-\log\tfrac34=\mathbf{0.288\ \text{nats}}$ . The guaranteed single-shot work is $W^{(1)}_{\mathrm{ext}}=k_BT\,D_0=0.288\,k_BT$ — strictly less than the average $0.347\,k_BT$ , because worst-case success on both populated levels is limited by their combined Gibbs weight.

Step 3. Min-divergence (formation). $D_\infty(\rho\|\gamma)=\log\max_i(p_i/\gamma_i)=\log\max(1,\,2)=\ln2=\mathbf{0.693\ \text{nats}}$ , set by level $1$ whose population is twice its Gibbs weight. Forming the state costs $W^{(1)}_{\mathrm{form}}=k_BT\,D_\infty=0.693\,k_BT$ .

Step 4. The gap. The ordering $D_0=0.288\le D=0.347\le D_\infty=0.693$ holds, and the single-shot irreversibility is $W^{(1)}_{\mathrm{form}}-W^{(1)}_{\mathrm{ext}}=k_BT(0.693-0.288)=0.405\,k_BT=k_BT\ln\tfrac32$ . You pay $0.693\,k_BT$ to build the state but recover only $0.288\,k_BT$ from it — a genuine $0.405\,k_BT$ loss that survives any protocol and vanishes only in the many-copy limit.

Physics of the portrait

The resource theory of athermality is where two centuries of thermodynamics meet quantum information: a free energy Gibbs wrote for gases, sharpened by Szilard and Landauer into a currency for information, and recast by Horodecki, Oppenheim, and Brandão as a monotone whose one second law becomes many.

Origins · 1 minnonequilibrium free energy and how this was discovered

nonequilibrium free energy

F_β(ρ) = ⟨H⟩ − k_B T·S(ρ), the athermality monotone · quantum thermodynamics as a resource theory

The free energy is one of the oldest quantities in physics and one of the last understood as an information measure. Rudolf Clausius in 1865 named entropy and cast the second law as $\mathrm{d}S\ge\delta Q/T$ ; Josiah Willard Gibbs in the 1870s introduced the ensemble $e^{-\beta H}/Z$ and the free energy $F=U-TS$ that decides which way an isothermal process runs — but only near equilibrium. The information turn came in 1929, when Leó Szilard analysed a one-molecule engine and found that a bit of information about which side the molecule sat on was worth $k_BT\ln2$ of work, forging the first link between knowing and doing. Rolf Landauer closed the loop in 1961: erasing one bit — forgetting, not learning — must dissipate at least $k_BT\ln2$ , making logical irreversibility a thermodynamic cost and rescuing the second law from Maxwell's demon. The synthesis arrived with Michał Horodecki and Jonathan Oppenheim in 2013, who defined thermal operations and proved that single-copy thermodynamics is governed not by one free energy but by a whole family, and with Brandão, Horodecki, Ng, Oppenheim, and Wehner in 2015, who showed the second law splinters into a continuum of Rényi constraints recovered by $F_\beta$ only in the thermodynamic limit. What Gibbs wrote as one number is now the leading member of an infinite tower.

History

1876

— Gibbs introduces the canonical ensemble e^(−βH)/Z and the free energy F = U − TS

1929

— Szilard shows one bit of information is worth k_B T ln 2 of work in his engine

1961

— Landauer proves erasing one bit costs at least k_B T ln 2, taming Maxwell's demon

2013

— Horodecki and Oppenheim define thermal operations and single-shot free energies

2015

— Brandão, Horodecki, Ng, Oppenheim and Wehner derive the family of second laws

Going further

Texts and reviews. Horodecki & Oppenheim, "Fundamental limitations for quantum and nanoscale thermodynamics," Nat. Commun. 4, 2059 (2013) · Brandão, Horodecki, Ng, Oppenheim, Wehner, "The second laws of quantum thermodynamics," PNAS 112, 3275 (2015) · Goold, Huber, Riera, del Rio, Skrzypczyk, "The role of quantum information in thermodynamics — a topical review," J. Phys. A 49, 143001 (2016) · Lostaglio, "An introductory review of the resource theory approach to thermodynamics," Rep. Prog. Phys. 82, 114001 (2019) · Wilde, Quantum Information Theory, 2nd ed. · Preskill, Lecture Notes on Quantum Computation, Ch. 10.

Provenance. Definitions and theorems follow Horodecki–Oppenheim (2013) and Brandão et al. (2015), with the reviews of Goold et al. (2016) and Lostaglio (2019) for context; relative-entropy monotonicity and data processing are from Wilde. A verifier should recompute: (1) $D(\rho\|\gamma)=\beta[F_\beta(\rho)-F_\beta(\gamma)]$ from $\log\gamma=-\beta H-\log Z$ ; (2) Gibbs invariance $\mathcal{E}(\gamma)=\gamma$ from $[U,H_S+H_B]=0$ ; (3) the qutrit numbers $D_0=0.288$ , $D=0.347$ , $D_\infty=0.693$ nats with gap $\ln\tfrac32$ ; (4) Landauer $k_BT\ln2=17.9\ \mathrm{meV}$ and $k_BT\ln3=28.4\ \mathrm{meV}$ at $300\ \mathrm{K}$ . All numerical statements are standard textbook values.

Prerequisites. Von Neumann Entropy · Quantum Relative Entropy · Quantum Channels and Completely Positive Trace-Preserving Maps · Data-Processing Inequality · Shannon Entropy and Classical Information Measures. Leads to. Strong Subadditivity of Quantum Entropy · Holevo Bound · Quantum Joint, Conditional Entropy and Mutual Information.

Ask the tutor: Ask me to walk through why the single second law of thermodynamics fractures into a whole Rényi family the moment you drop from many copies to one.

What exactly does it mean to call a heat bath the free resource?

In a resource theory you first declare what is available for nothing; everything else is a resource you must pay for. Here the free thing is one heat bath at a fixed inverse temperature $\beta$ : you may draw as many ancillas as you like in their thermal Gibbs states, couple them with any energy-conserving unitary, and discard them. What you may not do is inject work from a hidden source or bring in a colder or hotter bath. The Gibbs state is the one state you can always make freely, so any deviation from it is athermality: the currency the theory tracks.

Why does the single second law split into infinitely many at the quantum scale?

Because a single number can only summarize a spectrum when fluctuations wash out. With many i.i.d. copies the quantum asymptotic equipartition property flattens the eigenvalue distribution and every $D_\alpha(\rho\|\gamma)$ converges to $D(\rho\|\gamma)$ , so one free energy $F_\beta$ decides all conversions — the textbook law. For a single copy the full spectrum matters, and exact conversion $\rho\to\sigma$ (energy-incoherent states, with catalysts) requires $D_\alpha(\rho\|\gamma)\ge D_\alpha(\sigma\|\gamma)$ for every $\alpha\ge0$ — a continuum of independent constraints, not one.

Is athermality the same idea as entanglement, just for thermodynamics?

They are cousins with the same skeleton but different flesh. Both are resource theories: fix free operations (LOCC for entanglement, thermal operations here), identify the free states (product states there, Gibbs states here), and study conversion under monotones — the machinery of monotones, asymptotic rates, single-shot bounds, and catalysis transfers almost verbatim. But the physics differs: entanglement's free operations are constrained by locality, athermality's by energy conservation at a fixed temperature. The deep new feature athermality adds is time-translation covariance, so coherence becomes a second resource with no analogue in the basic entanglement story.

How can extracting work and erasing a bit be the same physics?

They are two directions of the same free-energy ledger. Work extraction spends athermality: you let an out-of-equilibrium state relax toward the Gibbs state and route the released $-\Delta F_\beta$ into a battery. Erasure is the reverse: resetting a bit lowers the system's entropy, which raises its free energy, and since thermal operations cannot raise $F_\beta$ for free, you must supply the increase as work — exactly $k_BT\ln2$ per bit when $H=0$ , which is Landauer's bound. Same monotone, opposite sign.

What is a catalyst, and why does it change the rules?

A catalyst is an ancilla you borrow, use, and return in exactly its original state — no net change, so it costs nothing by the theory's own bookkeeping. Remarkably, its mere presence can unlock otherwise forbidden transformations, because during the protocol it temporarily lends structure the system exploits. Catalysis is what makes the Rényi family not just necessary but the precise necessary-and-sufficient conditions for energy-incoherent conversion: with catalysts, $\rho\to\sigma$ is possible if and only if $F_\alpha(\rho)\ge F_\alpha(\sigma)$ for all $\alpha\ge0$ — the sharpest statement that the "true" second laws form a continuum.

Why can't a thermal operation ever create quantum coherence?

Because energy conservation is a symmetry, and symmetries constrain what maps can do. The condition $[U,H_S+H_B]=0$ makes every thermal operation commute with the time-translation the total Hamiltonian generates — the channel is covariant under $e^{-iHt}$ . A covariant channel cannot manufacture coherence between different-energy eigenspaces, just as a rotationally symmetric process cannot create net angular momentum: coherence is like a conserved charge the free operations must respect. Free energy, depending only on populations and entropy, is blind to it, so separate asymmetry monotones are needed — the genuinely quantum boundary of the theory.